There has been significant interest in "extreme" compression of large language models (LLMs), i.e., to 1-2 bits per parameter, which allows such models to be executed efficiently on resource-constrained devices. Existing work focused on improved one-shot quantization techniques and weight representations; yet, purely post-training approaches are reaching diminishing returns in terms of the accuracy-vs-bit-width trade-off. State-of-the-art quantization methods such as QuIP# and AQLM include fine-tuning (part of) the compressed parameters over a limited amount of calibration data; however, such fine-tuning techniques over compressed weights often make exclusive use of straight-through estimators (STE), whose performance is not well-understood in this setting. In this work, we question the use of STE for extreme LLM compression, showing that it can be sub-optimal, and perform a systematic study of quantization-aware fine-tuning strategies for LLMs. We propose PV-Tuning - a representation-agnostic framework that generalizes and improves upon existing fine-tuning strategies, and provides convergence guarantees in restricted cases. On the practical side, when used for 1-2 bit vector quantization, PV-Tuning outperforms prior techniques for highly-performant models such as Llama and Mistral. Using PV-Tuning, we achieve the first Pareto-optimal quantization for Llama 2 family models at 2 bits per parameter.
Modern large language models (LLMs) have established state-of-the-art performance through architectural improvements, but still require significant computational cost for inference. In an effort to reduce the inference cost, post-training quantization (PTQ) has become a popular approach, quantizing weights and activations to lower precision, such as INT8. In this paper, we reveal the challenges of activation quantization in GLU variants, which are widely used in feed-forward network (FFN) of modern LLMs, such as LLaMA family. The problem is that severe local quantization errors, caused by excessive magnitudes of activation in GLU variants, significantly degrade the performance of the quantized LLM. We denote these activations as activation spikes. Our further observations provide a systematic pattern of activation spikes: 1) The activation spikes occur in the FFN of specific layers, particularly in the early and late layers, 2) The activation spikes are dedicated to a couple of tokens, rather than being shared across a sequence. Based on our observations, we propose two empirical methods, Quantization-free Module (QFeM) and Quantization-free Prefix (QFeP), to isolate the activation spikes during quantization. Our extensive experiments validate the effectiveness of the proposed methods for the activation quantization, especially with coarse-grained scheme, of latest LLMs with GLU variants, including LLaMA-2/3, Mistral, Mixtral, SOLAR, and Gemma. In particular, our methods enhance the current alleviation techniques (e.g., SmoothQuant) that fail to control the activation spikes. Code is available at https://github.com/onnoo/activation-spikes.
Learning discrete representations with vector quantization (VQ) has emerged as a powerful approach in various generative models. However, most VQ-based models rely on a single, fixed-rate codebook, requiring extensive retraining for new bitrates or efficiency requirements. We introduce Rate-Adaptive Quantization (RAQ), a multi-rate codebook adaptation framework for VQ-based generative models. RAQ applies a data-driven approach to generate variable-rate codebooks from a single baseline VQ model, enabling flexible tradeoffs between compression and reconstruction fidelity. Additionally, we provide a simple clustering-based procedure for pre-trained VQ models, offering an alternative when retraining is infeasible. Our experiments show that RAQ performs effectively across multiple rates, often outperforming conventional fixed-rate VQ baselines. By enabling a single system to seamlessly handle diverse bitrate requirements, RAQ extends the adaptability of VQ-based generative models and broadens their applicability to data compression, reconstruction, and generation tasks.
Recent developments in large-scale pre-trained text-to-image diffusion models have significantly improved the generation of high-fidelity images, particularly with the emergence of diffusion models based on transformer architecture (DiTs). Among these diffusion models, diffusion transformers have demonstrated superior image generation capabilities, boosting lower FID scores and higher scalability. However, deploying large-scale DiT models can be expensive due to their extensive parameter numbers. Although existing research has explored efficient deployment techniques for diffusion models such as model quantization, there is still little work concerning DiT-based models. To tackle this research gap, in this paper, we propose TerDiT, a quantization-aware training (QAT) and efficient deployment scheme for ternary diffusion models with transformers. We focus on the ternarization of DiT networks and scale model sizes from 600M to 4.2B. Our work contributes to the exploration of efficient deployment strategies for large-scale DiT models, demonstrating the feasibility of training extremely low-bit diffusion transformer models from scratch while maintaining competitive image generation capacities compared to full-precision models. Code will be available at https://github.com/Lucky-Lance/TerDiT.
Unrolled deep neural networks have attracted significant attention for their success in various practical applications. In this paper, we explore an application of deep unrolling in the direction of arrival (DoA) estimation problem when coarse quantization is applied to the measurements. We present a compressed sensing formulation for DoA estimation from one-bit data in which estimating target DoAs requires recovering a sparse signal from a limited number of severely quantized linear measurements. In particular, we exploit covariance recovery from one-bit dither samples. To recover the covariance of transmitted signal, the learned iterative shrinkage and thresholding algorithm (LISTA) is employed fed by one-bit data. We demonstrate that the upper bound of estimation performance is governed by the recovery error of the transmitted signal covariance matrix. Through numerical experiments, we demonstrate the proposed LISTA-based algorithm's capability in estimating target locations. The code employed in this study is available online.
Key-value~(KV) caching is an important technique to accelerate the inference of large language models~(LLMs), but incurs significant memory overhead. To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment. In this paper, we introduce \textbf{DecoQuant}, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache. Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors. Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range. Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor. Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant. Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a $\sim$75\% reduction in memory footprint while maintaining comparable generation quality.
Searching for approximate nearest neighbors (ANN) in the high-dimensional Euclidean space is a pivotal problem. Recently, with the help of fast SIMD-based implementations, Product Quantization (PQ) and its variants can often efficiently and accurately estimate the distances between the vectors and have achieved great success in the in-memory ANN search. Despite their empirical success, we note that these methods do not have a theoretical error bound and are observed to fail disastrously on some real-world datasets. Motivated by this, we propose a new randomized quantization method named RaBitQ, which quantizes $D$-dimensional vectors into $D$-bit strings. RaBitQ guarantees a sharp theoretical error bound and provides good empirical accuracy at the same time. In addition, we introduce efficient implementations of RaBitQ, supporting to estimate the distances with bitwise operations or SIMD-based operations. Extensive experiments on real-world datasets confirm that (1) our method outperforms PQ and its variants in terms of accuracy-efficiency trade-off by a clear margin and (2) its empirical performance is well-aligned with our theoretical analysis.
Model quantization is widely used to compress and accelerate deep neural networks. However, recent studies have revealed the feasibility of weaponizing model quantization via implanting quantization-conditioned backdoors (QCBs). These special backdoors stay dormant on released full-precision models but will come into effect after standard quantization. Due to the peculiarity of QCBs, existing defenses have minor effects on reducing their threats or are even infeasible. In this paper, we conduct the first in-depth analysis of QCBs. We reveal that the activation of existing QCBs primarily stems from the nearest rounding operation and is closely related to the norms of neuron-wise truncation errors (i.e., the difference between the continuous full-precision weights and its quantized version). Motivated by these insights, we propose Error-guided Flipped Rounding with Activation Preservation (EFRAP), an effective and practical defense against QCBs. Specifically, EFRAP learns a non-nearest rounding strategy with neuron-wise error norm and layer-wise activation preservation guidance, flipping the rounding strategies of neurons crucial for backdoor effects but with minimal impact on clean accuracy. Extensive evaluations on benchmark datasets demonstrate that our EFRAP can defeat state-of-the-art QCB attacks under various settings. Code is available at https://github.com/AntigoneRandy/QuantBackdoor_EFRAP.
By harnessing the capabilities of large language models (LLMs), recent large multimodal models (LMMs) have shown remarkable versatility in open-world multimodal understanding. Nevertheless, they are usually parameter-heavy and computation-intensive, thus hindering their applicability in resource-constrained scenarios. To this end, several lightweight LMMs have been proposed successively to maximize the capabilities under constrained scale (e.g., 3B). Despite the encouraging results achieved by these methods, most of them only focus on one or two aspects of the design space, and the key design choices that influence model capability have not yet been thoroughly investigated. In this paper, we conduct a systematic study for lightweight LMMs from the aspects of model architecture, training strategy, and training data. Based on our findings, we obtain Imp -- a family of highly capable LMMs at the 2B-4B scales. Notably, our Imp-3B model steadily outperforms all the existing lightweight LMMs of similar size, and even surpasses the state-of-the-art LMMs at the 13B scale. With low-bit quantization and resolution reduction techniques, our Imp model can be deployed on a Qualcomm Snapdragon 8Gen3 mobile chip with a high inference speed of about 13 tokens/s.
Deep Neural Networks (DNNs) are known to be vulnerable to adversarial examples. Further, these adversarial examples are found to be transferable from the source network in which they are crafted to a black-box target network. As the trend of using deep learning on embedded devices grows, it becomes relevant to study the transferability properties of adversarial examples among compressed networks. In this paper, we consider quantization as a network compression technique and evaluate the performance of transfer-based attacks when the source and target networks are quantized at different bitwidths. We explore how algorithm specific properties affect transferability by considering various adversarial example generation algorithms. Furthermore, we examine transferability in a more realistic scenario where the source and target networks may differ in bitwidth and other model-related properties like capacity and architecture. We find that although quantization reduces transferability, certain attack types demonstrate an ability to enhance it. Additionally, the average transferability of adversarial examples among quantized versions of a network can be used to estimate the transferability to quantized target networks with varying capacity and architecture.
In this paper, we propose feature-based federated transfer learning as a novel approach to improve communication efficiency by reducing the uplink payload by multiple orders of magnitude compared to that of existing approaches in federated learning and federated transfer learning. Specifically, in the proposed feature-based federated learning, we design the extracted features and outputs to be uploaded instead of parameter updates. For this distributed learning model, we determine the required payload and provide comparisons with the existing schemes. Subsequently, we analyze the robustness of feature-based federated transfer learning against packet loss, data insufficiency, and quantization. Finally, we address privacy considerations by defining and analyzing label privacy leakage and feature privacy leakage, and investigating mitigating approaches. For all aforementioned analyses, we evaluate the performance of the proposed learning scheme via experiments on an image classification task and a natural language processing task to demonstrate its effectiveness.
Recent advancements in large language models (LLMs) are propelling us toward artificial general intelligence with their remarkable emergent abilities and reasoning capabilities. However, the substantial computational and memory requirements limit the widespread adoption. Quantization, a key compression technique, can effectively mitigate these demands by compressing and accelerating LLMs, albeit with potential risks to accuracy. Numerous studies have aimed to minimize the accuracy loss associated with quantization. However, their quantization configurations vary from each other and cannot be fairly compared. In this paper, we present LLMC, a plug-and-play compression toolkit, to fairly and systematically explore the impact of quantization. LLMC integrates dozens of algorithms, models, and hardwares, offering high extensibility from integer to floating-point quantization, from LLM to vision-language (VLM) model, from fixed-bit to mixed precision, and from quantization to sparsification. Powered by this versatile toolkit, our benchmark covers three key aspects: calibration data, algorithms (three strategies), and data formats, providing novel insights and detailed analyses for further research and practical guidance for users. Our toolkit is available at https://github.com/ModelTC/llmc.
Due to the rising privacy concerns on sensitive client data and trained models like Transformers, secure multi-party computation (MPC) techniques are employed to enable secure inference despite attendant overhead. Existing works attempt to reduce the overhead using more MPC-friendly non-linear function approximations. However, the integration of quantization widely used in plaintext inference into the MPC domain remains unclear. To bridge this gap, we propose the framework named Ditto to enable more efficient quantization-aware secure Transformer inference. Concretely, we first incorporate an MPC-friendly quantization into Transformer inference and employ a quantization-aware distillation procedure to maintain the model utility. Then, we propose novel MPC primitives to support the type conversions that are essential in quantization and implement the quantization-aware MPC execution of secure quantized inference. This approach significantly decreases both computation and communication overhead, leading to improvements in overall efficiency. We conduct extensive experiments on Bert and GPT2 models to evaluate the performance of Ditto. The results demonstrate that Ditto is about $3.14\sim 4.40\times$ faster than MPCFormer (ICLR 2023) and $1.44\sim 2.35\times$ faster than the state-of-the-art work PUMA with negligible utility degradation.
Motivated by the huge success of Transformers in the field of natural language processing (NLP), Vision Transformers (ViTs) have been rapidly developed and achieved remarkable performance in various computer vision tasks. However, their huge model sizes and intensive computations hinder ViTs' deployment on embedded devices, calling for effective model compression methods, such as quantization. Unfortunately, due to the existence of hardware-unfriendly and quantization-sensitive non-linear operations, particularly {Softmax}, it is non-trivial to completely quantize all operations in ViTs, yielding either significant accuracy drops or non-negligible hardware costs. In response to challenges associated with \textit{standard ViTs}, we focus our attention towards the quantization and acceleration for \textit{efficient ViTs}, which not only eliminate the troublesome Softmax but also integrate linear attention with low computational complexity, and propose Trio-ViT accordingly. Specifically, at the algorithm level, we develop a {tailored post-training quantization engine} taking the unique activation distributions of Softmax-free efficient ViTs into full consideration, aiming to boost quantization accuracy. Furthermore, at the hardware level, we build an accelerator dedicated to the specific Convolution-Transformer hybrid architecture of efficient ViTs, thereby enhancing hardware efficiency. Extensive experimental results consistently prove the effectiveness of our Trio-ViT framework. {Particularly, we can gain up to $\uparrow$$\mathbf{3.6}\times$, $\uparrow$$\mathbf{5.0}\times$, and $\uparrow$$\mathbf{7.3}\times$ FPS under comparable accuracy over state-of-the-art ViT accelerators, as well as $\uparrow$$\mathbf{6.0}\times$, $\uparrow$$\mathbf{1.5}\times$, and $\uparrow$$\mathbf{2.1}\times$ DSP efficiency.} Codes are available at \url{https://github.com/shihuihong214/Trio-ViT}.
Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/omniserve.
Segment Anything Model (SAM) has achieved impressive performance in many computer vision tasks. However, as a large-scale model, the immense memory and computation costs hinder its practical deployment. In this paper, we propose a post-training quantization (PTQ) framework for Segment Anything Model, namely PTQ4SAM. First, we investigate the inherent bottleneck of SAM quantization attributed to the bimodal distribution in post-Key-Linear activations. We analyze its characteristics from both per-tensor and per-channel perspectives, and propose a Bimodal Integration strategy, which utilizes a mathematically equivalent sign operation to transform the bimodal distribution into a relatively easy-quantized normal distribution offline. Second, SAM encompasses diverse attention mechanisms (i.e., self-attention and two-way cross-attention), resulting in substantial variations in the post-Softmax distributions. Therefore, we introduce an Adaptive Granularity Quantization for Softmax through searching the optimal power-of-two base, which is hardware-friendly. Extensive experimental results across various vision tasks (instance segmentation, semantic segmentation and object detection), datasets and model variants show the superiority of PTQ4SAM. For example, when quantizing SAM-L to 6-bit, we achieve lossless accuracy for instance segmentation, about 0.5\% drop with theoretical 3.9$\times$ acceleration. The code is available at \url{https://github.com/chengtao-lv/PTQ4SAM}.
The increasing size of large language models (LLMs) traditionally requires low-precision integer formats to meet strict latency and power demands. Yet recently, alternative formats such as Normal Float (NF4) have increased model accuracy at the cost of increased chip area. In this work, we first conduct a large-scale analysis of LLM weights and activations across 30 networks and conclude that most distributions follow a Student's t-distribution. We then derive a new theoretically optimal format, Student Float (SF4), that improves over NF4 across modern LLMs, for example increasing the average accuracy on LLaMA2-7B by 0.76% across tasks. Using this format as a high-accuracy reference, we then propose augmenting E2M1 with two variants of supernormal support for higher model accuracy. Finally, we explore the quality and efficiency frontier across 11 datatypes by evaluating their model accuracy and hardware complexity. We discover a Pareto curve composed of INT4, E2M1, and E2M1 with supernormal support, which offers a continuous tradeoff between model accuracy and chip area. For example, E2M1 with supernormal support increases the accuracy of Phi-2 by up to 2.19% with 1.22% area overhead, enabling more LLM-based applications to be run at four bits. The supporting code is hosted at https://github.com/cornell-zhang/llm-datatypes.
Text-driven human motion generation based on diffusion strategies establishes a reliable foundation for multimodal applications in human-computer interactions. However, existing advances face significant efficiency challenges due to the substantial computational overhead of iteratively solving for nonlinear reverse diffusion trajectories during the inference phase. To this end, we propose the motion latent consistency training framework (MLCT), which precomputes reverse diffusion trajectories from raw data in the training phase and enables few-step or single-step inference via self-consistency constraints in the inference phase. Specifically, a motion autoencoder with quantization constraints is first proposed for constructing concise and bounded solution distributions for motion diffusion processes. Subsequently, a classifier-free guidance format is constructed via an additional unconditional loss function to accomplish the precomputation of conditional diffusion trajectories in the training phase. Finally, a clustering guidance module based on the K-nearest-neighbor algorithm is developed for the chain-conduction optimization mechanism of self-consistency constraints, which provides additional references of solution distributions at a small query cost. By combining these enhancements, we achieve stable and consistency training in non-pixel modality and latent representation spaces. Benchmark experiments demonstrate that our method significantly outperforms traditional consistency distillation methods with reduced training cost and enhances the consistency model to perform comparably to state-of-the-art models with lower inference costs.
Vector quantization, renowned for its unparalleled feature compression capabilities, has been a prominent topic in signal processing and machine learning research for several decades and remains widely utilized today. With the emergence of large models and generative AI, vector quantization has gained popularity in recommender systems, establishing itself as a preferred solution. This paper starts with a comprehensive review of vector quantization techniques. It then explores systematic taxonomies of vector quantization methods for recommender systems (VQ4Rec), examining their applications from multiple perspectives. Further, it provides a thorough introduction to research efforts in diverse recommendation scenarios, including efficiency-oriented approaches and quality-oriented approaches. Finally, the survey analyzes the remaining challenges and anticipates future trends in VQ4Rec, including the challenges associated with the training of vector quantization, the opportunities presented by large language models, and emerging trends in multimodal recommender systems. We hope this survey can pave the way for future researchers in the recommendation community and accelerate their exploration in this promising field.
The development of model compression is continuously motivated by the evolution of various neural network accelerators with ASIC or FPGA. On the algorithm side, the ultimate goal of quantization or pruning is accelerating the expensive DNN computations on low-power hardware. However, such a "design-and-deploy" workflow faces under-explored challenges in the current hardware-algorithm co-design community. First, although the state-of-the-art quantization algorithm can achieve low precision with negligible degradation of accuracy, the latest deep learning framework (e.g., PyTorch) can only support non-customizable 8-bit precision, data format, and parameter extraction. Secondly, the objective of quantization is to enable the computation with low-precision data. However, the current SoTA algorithm treats the quantized integer as an intermediate result, while the final output of the quantizer is the "discretized" floating-point values, ignoring the practical needs and adding additional workload to hardware designers for integer parameter extraction and layer fusion. Finally, the compression toolkits designed by the industry are constrained to their in-house product or a handful of algorithms. The limited degree of freedom in the current toolkit and the under-explored customization hinder the prototype ASIC or FPGA-based accelerator design. To resolve these challenges, we propose Torch2Chip, an open-sourced, fully customizable, and high-performance toolkit that supports user-designed compression followed by automatic model fusion and parameter extraction. Torch2Chip incorporates the hierarchical design workflow, and the user-customized compression algorithm will be directly packed into the deployment-ready format for prototype chip verification with either CNN or vision transformer (ViT). The code is available at https://github.com/SeoLabCornell/torch2chip.